Congress Pushes to Sanction Chinese Firms Stealing American AI Models
New legislation targets foreign entities using distillation attacks and espionage to replicate US AI models, backed by a House Select Committee report detailing China's systematic campaign to acquire frontier AI capabilities.
The U.S. House of Representatives is advancing legislation to punish foreign entities — principally in China — that extract capabilities from American AI models through what the bill's sponsors call industrial-scale fraud. The Deterring American AI Model Theft Act of 2026 (H.R. 8283), introduced by Representative Bill Huizenga, would direct the federal government to identify and sanction companies using improper query-and-copy techniques to replicate the outputs of leading U.S. AI systems, in addition to traditional theft through espionage.
The push follows a bipartisan investigation by the House Select Committee on China, whose report — titled "Buy What It Can, Steal What It Must" — details how Chinese state and private actors have combined chip smuggling, industrial espionage, and large-scale model distillation attacks to build AI capabilities they could not develop domestically at pace. The report identifies the 2025 release of DeepSeek R1 as the first major public example of a distillation attack: a model trained extensively on outputs from frontier American AI systems to approximate their behavior at a fraction of the development cost.
The legal dimension of AI theft has already produced convictions in U.S. federal courts. In January 2026, a jury found former Google software engineer Linwei Ding guilty on seven counts of economic espionage and seven counts of trade secret theft, after prosecutors established that between 2022 and 2023 he had stolen more than 2,000 pages of Google's AI-related trade secrets. Intelligence assessments shared with Congress describe "foreign entities, principally based in China," conducting large-scale extraction of American AI technology — including through fake research accounts that systematically query commercial API endpoints to map model behavior.
The House Foreign Affairs Committee has advanced a broader slate of bipartisan export control bills alongside H.R. 8283 that would tighten restrictions on AI-related technology flowing to China and Russia. The urgency has intensified as the competitive gap narrows: the Stanford HAI AI Index 2026 found that U.S. and Chinese frontier models are now trading positions at the top of global benchmarks, with performance margins as thin as 2.7 percentage points — compared to a gap of roughly 17 points two years ago.
Critics of aggressive export controls caution that restricting model access could fragment the global AI research ecosystem without meaningfully slowing Chinese development, since the underlying scientific literature is largely published openly. The White House has so far held back on sweeping new curbs, creating tension with members of Congress who argue that voluntary restraint cedes the initiative to adversaries already operating without self-imposed limits. The outcome of the legislative push could determine whether AI model outputs are ultimately treated as a form of intellectual property with enforceable international protections, or whether distillation attacks remain a legal gray zone that frontier labs must defend against on their own.
